Datasetpublic

timiiowolabi/muta_tutor_quality_sample

Muta STEM SFT Quality Sample A train-ready, supervised fine-tuning data for Muta Tutor: verified STEM questions paired with answers in five tutoring styles that covers mathematics, physics, chemistry, biology, and integrated science. Snapshot Property Value Rows 300,350 JSONL shards 14 Decompressed data 776,449,831 bytes (~740 MiB) Hub data file 39,070,661 bytes (~37.3 MiB), Zstandard Language English Split Train Fingerprint… See the full description on the dataset page: https://huggingface.co/datasets/timiiowolabi/muta_tutor_quality_sample.

sourceHugging Faceotherupdated 6d agoView on Hugging Face

0likes37downloads

Dataset Card

Muta STEM SFT Quality Sample

A train-ready, supervised fine-tuning data for Muta Tutor: verified STEM questions paired with answers in five tutoring styles that covers mathematics, physics, chemistry, biology, and integrated science.

Snapshot

Property	Value
Rows	300,350
JSONL shards	14
Decompressed data	776,449,831 bytes (~740 MiB)
Hub data file	39,070,661 bytes (~37.3 MiB), Zstandard
Language	English
Split	Train
Fingerprint	`037edf28cccff62d90c23e2d6caf56b9998dea6928080f98af2a513f9f92910e`

Source	Rows
Muta verified STEM v2	280,000
DeepMind Mathematics	20,000
Private WAEC e-learning review	47
Private Cheetah WAEC review	303

Subject	Rows	Tutoring style	Rows
Mathematics	150,311	Worked solution	120,350
Physics	54,030	Exam marking scheme	60,000
Chemistry	42,006	Misconception correction	60,000
Biology	42,003	Socratic hint	30,000
Integrated science	12,000	Concise answer	30,000

Record format

Every row has prompt, completion, and two-message messages fields for SFT, together with answer, subject, topic, difficulty, pedagogy, curriculum, provenance, verification, contamination, and tokenization metadata. Use messages for conversation-formatted training.

python

from datasets import load_dataset

train = load_dataset(
    "timiiowolabi/muta_tutor_quality_sample",
    split="train",
)

manifest.json and selector-provenance/ bind the exact selection, source evidence, audit receipts, and original shard hashes. The Hub data file is the lossless concatenation of the 14 manifest shards in listed order; HUGGINGFACE_COMPRESSION.json records its hash and byte counts.